Question that has dogged me for a while.

[Rodney W. Grimes]

> Consider the following network configuration.
> 
> 
> Internet ------- Gateway/Firewall ---------- Inside network (including a
> web host)
>             70.16.10.1/28     192.168.0.0/24  
> 
> The address of the outside is FICTIONAL, by the way.
> 
> For policy reasons I do NOT want the gateway machine to actually have
> the host on it.  There may be a number of things running on there but
> for the instant moment let's assume a standard pedestrian web host on
> port 80.
> 
> I have DNS pointing at "webhost.domain" @ 70.16.10.1.
> 
> I have NAT on the gateway (NAT internal to the kernel), and a "hole
> punch" in there with redirect_port tcp 192.168.1.1:80 70.16.10.1:80 as
> pat of the nat configuration statement.
> 
> This works fine for anyone on the outside.  HOWEVER, anyone on the
> INTERNAL network cannot see the host.
> 
> My NAT configuration looks like this:
> 
> #
> # Now divert all inbound packets that should go through NAT. Since this
> is NAT
> # it can only match a packet that previously was NATted on the way out.
> #
>         ${fwcmd} add 6000 nat 100 ip4 from any to me recv ${oif}
> #
> # Check stateful rules; we want to go there directly if there is a match
> #
>         ${fwcmd} add 7000 check-state
> #
> # Now pick up all *outbound* packets that originated from an inside address
> # and put them through NAT.  We then have
> # a packet with a local source address and we can allow it to be sent.
> # Therefore, if the packet is outbound let it pass and be done with it.
> #
>         ${fwcmd} add 8000 nat 100 ip4 from 192.168.0.0/16 to any xmit ${oif}
> >>    ${fwcmd} add 8001 nat 100 ip4 from 192.168.0.0/16 to ${oip}
>         ${fwcmd} add 8009 deny log ip4 from 192.168.0.0/16 to any xmit
> ${oif}
>         ${fwcmd} add 8010 pass ip4 from ${onet} to any xmit ${oif}
> 
> Without the ">>" line I get nothing; the packets get to the gateway and
> disappear.
> 
> With the ">>" line I DO get the packets re-emitted on the internal
> interface HOWEVER there is no translation to the internal interface IP
> on the gateway box.  So what I see on the internal box is this:
> 
> 11:19:16.369634 IP 192.168.10.40.60924 > 192.168.10.100.11443: Flags
> [S], seq 292171178, win 8192, options [mss 1460,nop,wscale
> 8,nop,nop,sackOK], length 0
> 11:19:16.369662 IP 192.168.10.100.11443 > 192.168.10.40.60924: Flags
> [S.], seq 3088872007, ack 292171179, win 65535, options [mss
> 1460,nop,wscale 6,sackOK,eol], length 0
> 
> Which won't work because the internal box got and sent this:

What is the NAT command running at instance 100?
Does it have an -alias_address of inside IP of router?
Are you tryint to use the same Nat instance to do both
the global internet acess translation and this special
inside loop back translation?  If so that usually can
not be made to work.

> 11:19:16.369337 IP 192.168.10.40.60924 > 70.169.168.7.11443: Flags [S],
> seq 292171178, win 8192, options [mss 1460,nop,wscale 8,nop,nop,sackOK],
> length 0
> 11:19:16.369433 IP 192.168.10.40.60925 > 70.169.168.7.11443: Flags [S],
> seq 2666765817, win 8192, options [mss 1460,nop,wscale
> 8,nop,nop,sackOK], length 0
> >> 11:19:16.369502 IP 192.168.10.40.60924 > 192.168.10.100.11443: Flags
> [S], seq 292171178, win 8192, options [mss 1460,nop,wscale
> 8,nop,nop,sackOK], length 0
> >> 11:19:16.369511 IP 192.168.10.40.60925 > 192.168.10.100.11443: Flags
> [S], seq 2666765817, win 8192, options [mss 1460,nop,wscale
> 8,nop,nop,sackOK], length 0
> 
> But since the gateway emitted the packet back on the wire *without*
> remapping the source address (to itself) it doesn't match on the client
> box 'cause there's no way back for it.

Yep.

> 
> There has to be a solution to this somewhere and I'm obviously missing
> it..... :)
> 
> -- 
> Karl Denninger
> karl at denninger.net <mailto:karl at denninger.net>
> /The Market Ticker/
> /[S/MIME encrypted email preferred]/

This is the classical "Loopback Nat Problem" of accessing machines
behind a NAT device that does not do the proper NATTing and 
routing of these packets.   Many small consumer routers got this
wrong for years, just like most ipfw code I have seen.

Most of the consumber devices have been fixed.  The trickery is
you need to NAT packets coming from the inside destined for the
outside IP into the internal IP.  That internal box MUST route
back via the NAT device, so the NATTED addresses for these
packets must be the inside IP of the router.

Your code looks to get this mostly right, but I think you have
missed something someplace.  I dont use kernel nat, and it looks
like you do so you well have to adjust these a little.

inside_ip_router="192.168.10.40"
outside_ip_router="70.16.10.1"
inside_ip_webserver="192.168.10.100"

#natd-vmxZ.conf should just be an empty file for this type of nat
/sbin/natd -f /etc/firewall/natd-vmxZ.conf  -port 8888 -alias_address ${192.168.10.40} 

Something like
# This takes inside traffic to outside port 80 address and
# remaps it to inside IP of router to send to web server
${fwcmd} add X divert 8888 tcp from 192.168.0.0/24 to ${outside_ip_router} port 80
# This translates the returning packets
${fwcmd} add X divert 8888 tcp from 192.168.0.0/24 port 80 to ${inside_ip_router}

-- 
Rod Grimes                                                 rgrimes at freebsd.org